Probabilistic Graphical Models for Semi-Supervised Traffic Classification
Traffic classification using machine learning\ncontinues to be an active research area. The\nmajority of work in this area uses off-the-shelf\nmachine learning tools and treats them as black-box\nclassifiers. This approach turns all the modelling\ncomplexity into a feature selection problem. In this\npaper, we build a problem-specific solution to the\ntraffic classification problem by designing a custom\nprobabilistic graphical model. Graphical models are\na modular framework to design classifiers which\nincorporate domain-specific knowledge. More\nspecifically, our solution introduces\nsemi-supervised learning which means we learn from\nboth labelled and unlabelled traffic flows. We show\nthat our solution performs competitively compared to\nprevious approaches while using less data and\nsimpler features.